Debian Codegraph @

src/Futhark/Pass/ExtractKernels/StreamKernel.hs

raw
1 {-# LANGUAGE TypeFamilies #-}
2
3 module Futhark.Pass.ExtractKernels.StreamKernel
4 ( segThreadCapped,
5 )
6 where
7
8 import Control.Monad
9 import Data.List ()
10 import Futhark.Analysis.PrimExp
11 import Futhark.IR
12 import Futhark.IR.GPU hiding
13 ( BasicOp,
14 Body,
15 Exp,
16 FParam,
17 FunDef,
18 LParam,
19 Lambda,
20 Pat,
21 PatElem,
22 Prog,
23 RetType,
24 Stm,
25 )
26 import Futhark.MonadFreshNames
27 import Futhark.Pass.ExtractKernels.BlockedKernel
28 import Futhark.Pass.ExtractKernels.ToGPU
29 import Futhark.Tools
30 import Prelude hiding (quot)
31
32 data KernelSize = KernelSize
33 { -- | Int64
34 kernelElementsPerThread :: SubExp,
35 -- | Int32
36 kernelNumThreads :: SubExp
37 }
38 deriving (Eq, Ord, Show)
39
40 numberOfBlocks ::
41 (MonadBuilder m, Op (Rep m) ~ HostOp inner (Rep m)) =>
42 String ->
43 SubExp ->
44 SubExp ->
45 m (SubExp, SubExp)
46 numberOfBlocks desc w tblock_size = do
47 max_num_tblocks_key <- nameFromString . prettyString <$> newVName (desc ++ "_num_tblocks")
48 num_tblocks <-
49 letSubExp "num_tblocks" $
50 Op $
51 SizeOp $
52 CalcNumBlocks w max_num_tblocks_key tblock_size
53 num_threads <-
54 letSubExp "num_threads" $
55 BasicOp $
56 BinOp (Mul Int64 OverflowUndef) num_tblocks tblock_size
57 pure (num_tblocks, num_threads)
58
59 -- | Like 'segThread', but cap the thread count to the input size.
60 -- This is more efficient for small kernels, e.g. summing a small
61 -- array.
62 segThreadCapped :: (MonadFreshNames m) => MkSegLevel GPU m
63 segThreadCapped ws desc r = do
64 w <-
65 letSubExp "nest_size"
66 =<< foldBinOp (Mul Int64 OverflowUndef) (intConst Int64 1) ws
67 tblock_size <- getSize (desc ++ "_tblock_size") SizeThreadBlock
68
69 case r of
70 ManyThreads -> do
71 usable_groups <-
72 letSubExp "segmap_usable_groups"
73 =<< eBinOp
74 (SDivUp Int64 Unsafe)
75 (eSubExp w)
76 (eSubExp =<< asIntS Int64 tblock_size)
77 let grid = KernelGrid (Count usable_groups) (Count tblock_size)
78 pure $ SegThread SegNoVirt (Just grid)
79 NoRecommendation v -> do
80 (num_tblocks, _) <- numberOfBlocks desc w tblock_size
81 let grid = KernelGrid (Count num_tblocks) (Count tblock_size)
82 pure $ SegThread v (Just grid)